Corey James- 294,033of 294,362 code chunks covered by the computed graph
- 0models involved in a dependency trace; it is a bounded database traversal
- 3indexes, code, past decisions and reference documents, behind one retrieval surface
Coding agents forget everything between sessions, rediscover the same facts about a codebase daily, and explain call paths that do not exist with complete confidence.
I built the memory and code-intelligence service every agent in my estate connects to. Dependency tracing is a graph query, and the code explainer narrates only edges the graph has proven and refuses when coverage is incomplete. Start with A code explainer that is not allowed to invent.
The anchor, a bounded traversal and the code on the path (in green) come from the parser and SQL; the model then narrates exactly that path. An edge enters the graph only when both its ends are indexed, and graph generations keep a half-built graph from being served.
The hollow node is a file the graph does not cover: a query that touches it returns coverage incomplete, with no free-form fallback.
A refusal beats a plausible guess, because the caveat is the first thing a reader stops seeing.
Session memory captures every tool call, extracts what matters, and re-injects it later.
Very long sessions are compressed rather than truncated. The same observation search reaches stored decisions and fixes, so a settled question is found before it is re-debugged.